1 // SPDX-License-Identifier: GPL-2.0 OR BSD-3-Clause 2 /* Copyright (c) 2021, Microsoft Corporation. */ 3 4 #include <linux/inetdevice.h> 5 #include <linux/etherdevice.h> 6 #include <linux/ethtool.h> 7 #include <linux/mm.h> 8 9 #include <net/checksum.h> 10 #include <net/ip6_checksum.h> 11 12 #include "mana.h" 13 14 /* Microsoft Azure Network Adapter (MANA) functions */ 15 16 static int mana_open(struct net_device *ndev) 17 { 18 struct mana_port_context *apc = netdev_priv(ndev); 19 int err; 20 21 err = mana_alloc_queues(ndev); 22 if (err) 23 return err; 24 25 apc->port_is_up = true; 26 27 /* Ensure port state updated before txq state */ 28 smp_wmb(); 29 30 netif_carrier_on(ndev); 31 netif_tx_wake_all_queues(ndev); 32 33 return 0; 34 } 35 36 static int mana_close(struct net_device *ndev) 37 { 38 struct mana_port_context *apc = netdev_priv(ndev); 39 40 if (!apc->port_is_up) 41 return 0; 42 43 return mana_detach(ndev, true); 44 } 45 46 static bool mana_can_tx(struct gdma_queue *wq) 47 { 48 return mana_gd_wq_avail_space(wq) >= MAX_TX_WQE_SIZE; 49 } 50 51 static unsigned int mana_checksum_info(struct sk_buff *skb) 52 { 53 if (skb->protocol == htons(ETH_P_IP)) { 54 struct iphdr *ip = ip_hdr(skb); 55 56 if (ip->protocol == IPPROTO_TCP) 57 return IPPROTO_TCP; 58 59 if (ip->protocol == IPPROTO_UDP) 60 return IPPROTO_UDP; 61 } else if (skb->protocol == htons(ETH_P_IPV6)) { 62 struct ipv6hdr *ip6 = ipv6_hdr(skb); 63 64 if (ip6->nexthdr == IPPROTO_TCP) 65 return IPPROTO_TCP; 66 67 if (ip6->nexthdr == IPPROTO_UDP) 68 return IPPROTO_UDP; 69 } 70 71 /* No csum offloading */ 72 return 0; 73 } 74 75 static int mana_map_skb(struct sk_buff *skb, struct mana_port_context *apc, 76 struct mana_tx_package *tp) 77 { 78 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 79 struct gdma_dev *gd = apc->ac->gdma_dev; 80 struct gdma_context *gc; 81 struct device *dev; 82 skb_frag_t *frag; 83 dma_addr_t da; 84 int i; 85 86 gc = gd->gdma_context; 87 dev = gc->dev; 88 da = dma_map_single(dev, skb->data, skb_headlen(skb), DMA_TO_DEVICE); 89 90 if (dma_mapping_error(dev, da)) 91 return -ENOMEM; 92 93 ash->dma_handle[0] = da; 94 ash->size[0] = skb_headlen(skb); 95 96 tp->wqe_req.sgl[0].address = ash->dma_handle[0]; 97 tp->wqe_req.sgl[0].mem_key = gd->gpa_mkey; 98 tp->wqe_req.sgl[0].size = ash->size[0]; 99 100 for (i = 0; i < skb_shinfo(skb)->nr_frags; i++) { 101 frag = &skb_shinfo(skb)->frags[i]; 102 da = skb_frag_dma_map(dev, frag, 0, skb_frag_size(frag), 103 DMA_TO_DEVICE); 104 105 if (dma_mapping_error(dev, da)) 106 goto frag_err; 107 108 ash->dma_handle[i + 1] = da; 109 ash->size[i + 1] = skb_frag_size(frag); 110 111 tp->wqe_req.sgl[i + 1].address = ash->dma_handle[i + 1]; 112 tp->wqe_req.sgl[i + 1].mem_key = gd->gpa_mkey; 113 tp->wqe_req.sgl[i + 1].size = ash->size[i + 1]; 114 } 115 116 return 0; 117 118 frag_err: 119 for (i = i - 1; i >= 0; i--) 120 dma_unmap_page(dev, ash->dma_handle[i + 1], ash->size[i + 1], 121 DMA_TO_DEVICE); 122 123 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 124 125 return -ENOMEM; 126 } 127 128 static int mana_start_xmit(struct sk_buff *skb, struct net_device *ndev) 129 { 130 enum mana_tx_pkt_format pkt_fmt = MANA_SHORT_PKT_FMT; 131 struct mana_port_context *apc = netdev_priv(ndev); 132 u16 txq_idx = skb_get_queue_mapping(skb); 133 struct gdma_dev *gd = apc->ac->gdma_dev; 134 bool ipv4 = false, ipv6 = false; 135 struct mana_tx_package pkg = {}; 136 struct netdev_queue *net_txq; 137 struct mana_stats *tx_stats; 138 struct gdma_queue *gdma_sq; 139 unsigned int csum_type; 140 struct mana_txq *txq; 141 struct mana_cq *cq; 142 int err, len; 143 144 if (unlikely(!apc->port_is_up)) 145 goto tx_drop; 146 147 if (skb_cow_head(skb, MANA_HEADROOM)) 148 goto tx_drop_count; 149 150 txq = &apc->tx_qp[txq_idx].txq; 151 gdma_sq = txq->gdma_sq; 152 cq = &apc->tx_qp[txq_idx].tx_cq; 153 154 pkg.tx_oob.s_oob.vcq_num = cq->gdma_id; 155 pkg.tx_oob.s_oob.vsq_frame = txq->vsq_frame; 156 157 if (txq->vp_offset > MANA_SHORT_VPORT_OFFSET_MAX) { 158 pkg.tx_oob.l_oob.long_vp_offset = txq->vp_offset; 159 pkt_fmt = MANA_LONG_PKT_FMT; 160 } else { 161 pkg.tx_oob.s_oob.short_vp_offset = txq->vp_offset; 162 } 163 164 pkg.tx_oob.s_oob.pkt_fmt = pkt_fmt; 165 166 if (pkt_fmt == MANA_SHORT_PKT_FMT) 167 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_short_oob); 168 else 169 pkg.wqe_req.inline_oob_size = sizeof(struct mana_tx_oob); 170 171 pkg.wqe_req.inline_oob_data = &pkg.tx_oob; 172 pkg.wqe_req.flags = 0; 173 pkg.wqe_req.client_data_unit = 0; 174 175 pkg.wqe_req.num_sge = 1 + skb_shinfo(skb)->nr_frags; 176 WARN_ON_ONCE(pkg.wqe_req.num_sge > 30); 177 178 if (pkg.wqe_req.num_sge <= ARRAY_SIZE(pkg.sgl_array)) { 179 pkg.wqe_req.sgl = pkg.sgl_array; 180 } else { 181 pkg.sgl_ptr = kmalloc_array(pkg.wqe_req.num_sge, 182 sizeof(struct gdma_sge), 183 GFP_ATOMIC); 184 if (!pkg.sgl_ptr) 185 goto tx_drop_count; 186 187 pkg.wqe_req.sgl = pkg.sgl_ptr; 188 } 189 190 if (skb->protocol == htons(ETH_P_IP)) 191 ipv4 = true; 192 else if (skb->protocol == htons(ETH_P_IPV6)) 193 ipv6 = true; 194 195 if (skb_is_gso(skb)) { 196 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 197 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 198 199 pkg.tx_oob.s_oob.comp_iphdr_csum = 1; 200 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 201 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 202 203 pkg.wqe_req.client_data_unit = skb_shinfo(skb)->gso_size; 204 pkg.wqe_req.flags = GDMA_WR_OOB_IN_SGL | GDMA_WR_PAD_BY_SGE0; 205 if (ipv4) { 206 ip_hdr(skb)->tot_len = 0; 207 ip_hdr(skb)->check = 0; 208 tcp_hdr(skb)->check = 209 ~csum_tcpudp_magic(ip_hdr(skb)->saddr, 210 ip_hdr(skb)->daddr, 0, 211 IPPROTO_TCP, 0); 212 } else { 213 ipv6_hdr(skb)->payload_len = 0; 214 tcp_hdr(skb)->check = 215 ~csum_ipv6_magic(&ipv6_hdr(skb)->saddr, 216 &ipv6_hdr(skb)->daddr, 0, 217 IPPROTO_TCP, 0); 218 } 219 } else if (skb->ip_summed == CHECKSUM_PARTIAL) { 220 csum_type = mana_checksum_info(skb); 221 222 if (csum_type == IPPROTO_TCP) { 223 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 224 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 225 226 pkg.tx_oob.s_oob.comp_tcp_csum = 1; 227 pkg.tx_oob.s_oob.trans_off = skb_transport_offset(skb); 228 229 } else if (csum_type == IPPROTO_UDP) { 230 pkg.tx_oob.s_oob.is_outer_ipv4 = ipv4; 231 pkg.tx_oob.s_oob.is_outer_ipv6 = ipv6; 232 233 pkg.tx_oob.s_oob.comp_udp_csum = 1; 234 } else { 235 /* Can't do offload of this type of checksum */ 236 if (skb_checksum_help(skb)) 237 goto free_sgl_ptr; 238 } 239 } 240 241 if (mana_map_skb(skb, apc, &pkg)) 242 goto free_sgl_ptr; 243 244 skb_queue_tail(&txq->pending_skbs, skb); 245 246 len = skb->len; 247 net_txq = netdev_get_tx_queue(ndev, txq_idx); 248 249 err = mana_gd_post_work_request(gdma_sq, &pkg.wqe_req, 250 (struct gdma_posted_wqe_info *)skb->cb); 251 if (!mana_can_tx(gdma_sq)) { 252 netif_tx_stop_queue(net_txq); 253 apc->eth_stats.stop_queue++; 254 } 255 256 if (err) { 257 (void)skb_dequeue_tail(&txq->pending_skbs); 258 netdev_warn(ndev, "Failed to post TX OOB: %d\n", err); 259 err = NETDEV_TX_BUSY; 260 goto tx_busy; 261 } 262 263 err = NETDEV_TX_OK; 264 atomic_inc(&txq->pending_sends); 265 266 mana_gd_wq_ring_doorbell(gd->gdma_context, gdma_sq); 267 268 /* skb may be freed after mana_gd_post_work_request. Do not use it. */ 269 skb = NULL; 270 271 tx_stats = &txq->stats; 272 u64_stats_update_begin(&tx_stats->syncp); 273 tx_stats->packets++; 274 tx_stats->bytes += len; 275 u64_stats_update_end(&tx_stats->syncp); 276 277 tx_busy: 278 if (netif_tx_queue_stopped(net_txq) && mana_can_tx(gdma_sq)) { 279 netif_tx_wake_queue(net_txq); 280 apc->eth_stats.wake_queue++; 281 } 282 283 kfree(pkg.sgl_ptr); 284 return err; 285 286 free_sgl_ptr: 287 kfree(pkg.sgl_ptr); 288 tx_drop_count: 289 ndev->stats.tx_dropped++; 290 tx_drop: 291 dev_kfree_skb_any(skb); 292 return NETDEV_TX_OK; 293 } 294 295 static void mana_get_stats64(struct net_device *ndev, 296 struct rtnl_link_stats64 *st) 297 { 298 struct mana_port_context *apc = netdev_priv(ndev); 299 unsigned int num_queues = apc->num_queues; 300 struct mana_stats *stats; 301 unsigned int start; 302 u64 packets, bytes; 303 int q; 304 305 if (!apc->port_is_up) 306 return; 307 308 netdev_stats_to_stats64(st, &ndev->stats); 309 310 for (q = 0; q < num_queues; q++) { 311 stats = &apc->rxqs[q]->stats; 312 313 do { 314 start = u64_stats_fetch_begin_irq(&stats->syncp); 315 packets = stats->packets; 316 bytes = stats->bytes; 317 } while (u64_stats_fetch_retry_irq(&stats->syncp, start)); 318 319 st->rx_packets += packets; 320 st->rx_bytes += bytes; 321 } 322 323 for (q = 0; q < num_queues; q++) { 324 stats = &apc->tx_qp[q].txq.stats; 325 326 do { 327 start = u64_stats_fetch_begin_irq(&stats->syncp); 328 packets = stats->packets; 329 bytes = stats->bytes; 330 } while (u64_stats_fetch_retry_irq(&stats->syncp, start)); 331 332 st->tx_packets += packets; 333 st->tx_bytes += bytes; 334 } 335 } 336 337 static int mana_get_tx_queue(struct net_device *ndev, struct sk_buff *skb, 338 int old_q) 339 { 340 struct mana_port_context *apc = netdev_priv(ndev); 341 u32 hash = skb_get_hash(skb); 342 struct sock *sk = skb->sk; 343 int txq; 344 345 txq = apc->indir_table[hash & MANA_INDIRECT_TABLE_MASK]; 346 347 if (txq != old_q && sk && sk_fullsock(sk) && 348 rcu_access_pointer(sk->sk_dst_cache)) 349 sk_tx_queue_set(sk, txq); 350 351 return txq; 352 } 353 354 static u16 mana_select_queue(struct net_device *ndev, struct sk_buff *skb, 355 struct net_device *sb_dev) 356 { 357 int txq; 358 359 if (ndev->real_num_tx_queues == 1) 360 return 0; 361 362 txq = sk_tx_queue_get(skb->sk); 363 364 if (txq < 0 || skb->ooo_okay || txq >= ndev->real_num_tx_queues) { 365 if (skb_rx_queue_recorded(skb)) 366 txq = skb_get_rx_queue(skb); 367 else 368 txq = mana_get_tx_queue(ndev, skb, txq); 369 } 370 371 return txq; 372 } 373 374 static const struct net_device_ops mana_devops = { 375 .ndo_open = mana_open, 376 .ndo_stop = mana_close, 377 .ndo_select_queue = mana_select_queue, 378 .ndo_start_xmit = mana_start_xmit, 379 .ndo_validate_addr = eth_validate_addr, 380 .ndo_get_stats64 = mana_get_stats64, 381 }; 382 383 static void mana_cleanup_port_context(struct mana_port_context *apc) 384 { 385 kfree(apc->rxqs); 386 apc->rxqs = NULL; 387 } 388 389 static int mana_init_port_context(struct mana_port_context *apc) 390 { 391 apc->rxqs = kcalloc(apc->num_queues, sizeof(struct mana_rxq *), 392 GFP_KERNEL); 393 394 return !apc->rxqs ? -ENOMEM : 0; 395 } 396 397 static int mana_send_request(struct mana_context *ac, void *in_buf, 398 u32 in_len, void *out_buf, u32 out_len) 399 { 400 struct gdma_context *gc = ac->gdma_dev->gdma_context; 401 struct gdma_resp_hdr *resp = out_buf; 402 struct gdma_req_hdr *req = in_buf; 403 struct device *dev = gc->dev; 404 static atomic_t activity_id; 405 int err; 406 407 req->dev_id = gc->mana.dev_id; 408 req->activity_id = atomic_inc_return(&activity_id); 409 410 err = mana_gd_send_request(gc, in_len, in_buf, out_len, 411 out_buf); 412 if (err || resp->status) { 413 dev_err(dev, "Failed to send mana message: %d, 0x%x\n", 414 err, resp->status); 415 return err ? err : -EPROTO; 416 } 417 418 if (req->dev_id.as_uint32 != resp->dev_id.as_uint32 || 419 req->activity_id != resp->activity_id) { 420 dev_err(dev, "Unexpected mana message response: %x,%x,%x,%x\n", 421 req->dev_id.as_uint32, resp->dev_id.as_uint32, 422 req->activity_id, resp->activity_id); 423 return -EPROTO; 424 } 425 426 return 0; 427 } 428 429 static int mana_verify_resp_hdr(const struct gdma_resp_hdr *resp_hdr, 430 const enum mana_command_code expected_code, 431 const u32 min_size) 432 { 433 if (resp_hdr->response.msg_type != expected_code) 434 return -EPROTO; 435 436 if (resp_hdr->response.msg_version < GDMA_MESSAGE_V1) 437 return -EPROTO; 438 439 if (resp_hdr->response.msg_size < min_size) 440 return -EPROTO; 441 442 return 0; 443 } 444 445 static int mana_query_device_cfg(struct mana_context *ac, u32 proto_major_ver, 446 u32 proto_minor_ver, u32 proto_micro_ver, 447 u16 *max_num_vports) 448 { 449 struct gdma_context *gc = ac->gdma_dev->gdma_context; 450 struct mana_query_device_cfg_resp resp = {}; 451 struct mana_query_device_cfg_req req = {}; 452 struct device *dev = gc->dev; 453 int err = 0; 454 455 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_DEV_CONFIG, 456 sizeof(req), sizeof(resp)); 457 req.proto_major_ver = proto_major_ver; 458 req.proto_minor_ver = proto_minor_ver; 459 req.proto_micro_ver = proto_micro_ver; 460 461 err = mana_send_request(ac, &req, sizeof(req), &resp, sizeof(resp)); 462 if (err) { 463 dev_err(dev, "Failed to query config: %d", err); 464 return err; 465 } 466 467 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_DEV_CONFIG, 468 sizeof(resp)); 469 if (err || resp.hdr.status) { 470 dev_err(dev, "Invalid query result: %d, 0x%x\n", err, 471 resp.hdr.status); 472 if (!err) 473 err = -EPROTO; 474 return err; 475 } 476 477 *max_num_vports = resp.max_num_vports; 478 479 return 0; 480 } 481 482 static int mana_query_vport_cfg(struct mana_port_context *apc, u32 vport_index, 483 u32 *max_sq, u32 *max_rq, u32 *num_indir_entry) 484 { 485 struct mana_query_vport_cfg_resp resp = {}; 486 struct mana_query_vport_cfg_req req = {}; 487 int err; 488 489 mana_gd_init_req_hdr(&req.hdr, MANA_QUERY_VPORT_CONFIG, 490 sizeof(req), sizeof(resp)); 491 492 req.vport_index = vport_index; 493 494 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 495 sizeof(resp)); 496 if (err) 497 return err; 498 499 err = mana_verify_resp_hdr(&resp.hdr, MANA_QUERY_VPORT_CONFIG, 500 sizeof(resp)); 501 if (err) 502 return err; 503 504 if (resp.hdr.status) 505 return -EPROTO; 506 507 *max_sq = resp.max_num_sq; 508 *max_rq = resp.max_num_rq; 509 *num_indir_entry = resp.num_indirection_ent; 510 511 apc->port_handle = resp.vport; 512 ether_addr_copy(apc->mac_addr, resp.mac_addr); 513 514 return 0; 515 } 516 517 static int mana_cfg_vport(struct mana_port_context *apc, u32 protection_dom_id, 518 u32 doorbell_pg_id) 519 { 520 struct mana_config_vport_resp resp = {}; 521 struct mana_config_vport_req req = {}; 522 int err; 523 524 mana_gd_init_req_hdr(&req.hdr, MANA_CONFIG_VPORT_TX, 525 sizeof(req), sizeof(resp)); 526 req.vport = apc->port_handle; 527 req.pdid = protection_dom_id; 528 req.doorbell_pageid = doorbell_pg_id; 529 530 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 531 sizeof(resp)); 532 if (err) { 533 netdev_err(apc->ndev, "Failed to configure vPort: %d\n", err); 534 goto out; 535 } 536 537 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_TX, 538 sizeof(resp)); 539 if (err || resp.hdr.status) { 540 netdev_err(apc->ndev, "Failed to configure vPort: %d, 0x%x\n", 541 err, resp.hdr.status); 542 if (!err) 543 err = -EPROTO; 544 545 goto out; 546 } 547 548 apc->tx_shortform_allowed = resp.short_form_allowed; 549 apc->tx_vp_offset = resp.tx_vport_offset; 550 out: 551 return err; 552 } 553 554 static int mana_cfg_vport_steering(struct mana_port_context *apc, 555 enum TRI_STATE rx, 556 bool update_default_rxobj, bool update_key, 557 bool update_tab) 558 { 559 u16 num_entries = MANA_INDIRECT_TABLE_SIZE; 560 struct mana_cfg_rx_steer_req *req = NULL; 561 struct mana_cfg_rx_steer_resp resp = {}; 562 struct net_device *ndev = apc->ndev; 563 mana_handle_t *req_indir_tab; 564 u32 req_buf_size; 565 int err; 566 567 req_buf_size = sizeof(*req) + sizeof(mana_handle_t) * num_entries; 568 req = kzalloc(req_buf_size, GFP_KERNEL); 569 if (!req) 570 return -ENOMEM; 571 572 mana_gd_init_req_hdr(&req->hdr, MANA_CONFIG_VPORT_RX, req_buf_size, 573 sizeof(resp)); 574 575 req->vport = apc->port_handle; 576 req->num_indir_entries = num_entries; 577 req->indir_tab_offset = sizeof(*req); 578 req->rx_enable = rx; 579 req->rss_enable = apc->rss_state; 580 req->update_default_rxobj = update_default_rxobj; 581 req->update_hashkey = update_key; 582 req->update_indir_tab = update_tab; 583 req->default_rxobj = apc->default_rxobj; 584 585 if (update_key) 586 memcpy(&req->hashkey, apc->hashkey, MANA_HASH_KEY_SIZE); 587 588 if (update_tab) { 589 req_indir_tab = (mana_handle_t *)(req + 1); 590 memcpy(req_indir_tab, apc->rxobj_table, 591 req->num_indir_entries * sizeof(mana_handle_t)); 592 } 593 594 err = mana_send_request(apc->ac, req, req_buf_size, &resp, 595 sizeof(resp)); 596 if (err) { 597 netdev_err(ndev, "Failed to configure vPort RX: %d\n", err); 598 goto out; 599 } 600 601 err = mana_verify_resp_hdr(&resp.hdr, MANA_CONFIG_VPORT_RX, 602 sizeof(resp)); 603 if (err) { 604 netdev_err(ndev, "vPort RX configuration failed: %d\n", err); 605 goto out; 606 } 607 608 if (resp.hdr.status) { 609 netdev_err(ndev, "vPort RX configuration failed: 0x%x\n", 610 resp.hdr.status); 611 err = -EPROTO; 612 } 613 out: 614 kfree(req); 615 return err; 616 } 617 618 static int mana_create_wq_obj(struct mana_port_context *apc, 619 mana_handle_t vport, 620 u32 wq_type, struct mana_obj_spec *wq_spec, 621 struct mana_obj_spec *cq_spec, 622 mana_handle_t *wq_obj) 623 { 624 struct mana_create_wqobj_resp resp = {}; 625 struct mana_create_wqobj_req req = {}; 626 struct net_device *ndev = apc->ndev; 627 int err; 628 629 mana_gd_init_req_hdr(&req.hdr, MANA_CREATE_WQ_OBJ, 630 sizeof(req), sizeof(resp)); 631 req.vport = vport; 632 req.wq_type = wq_type; 633 req.wq_gdma_region = wq_spec->gdma_region; 634 req.cq_gdma_region = cq_spec->gdma_region; 635 req.wq_size = wq_spec->queue_size; 636 req.cq_size = cq_spec->queue_size; 637 req.cq_moderation_ctx_id = cq_spec->modr_ctx_id; 638 req.cq_parent_qid = cq_spec->attached_eq; 639 640 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 641 sizeof(resp)); 642 if (err) { 643 netdev_err(ndev, "Failed to create WQ object: %d\n", err); 644 goto out; 645 } 646 647 err = mana_verify_resp_hdr(&resp.hdr, MANA_CREATE_WQ_OBJ, 648 sizeof(resp)); 649 if (err || resp.hdr.status) { 650 netdev_err(ndev, "Failed to create WQ object: %d, 0x%x\n", err, 651 resp.hdr.status); 652 if (!err) 653 err = -EPROTO; 654 goto out; 655 } 656 657 if (resp.wq_obj == INVALID_MANA_HANDLE) { 658 netdev_err(ndev, "Got an invalid WQ object handle\n"); 659 err = -EPROTO; 660 goto out; 661 } 662 663 *wq_obj = resp.wq_obj; 664 wq_spec->queue_index = resp.wq_id; 665 cq_spec->queue_index = resp.cq_id; 666 667 return 0; 668 out: 669 return err; 670 } 671 672 static void mana_destroy_wq_obj(struct mana_port_context *apc, u32 wq_type, 673 mana_handle_t wq_obj) 674 { 675 struct mana_destroy_wqobj_resp resp = {}; 676 struct mana_destroy_wqobj_req req = {}; 677 struct net_device *ndev = apc->ndev; 678 int err; 679 680 mana_gd_init_req_hdr(&req.hdr, MANA_DESTROY_WQ_OBJ, 681 sizeof(req), sizeof(resp)); 682 req.wq_type = wq_type; 683 req.wq_obj_handle = wq_obj; 684 685 err = mana_send_request(apc->ac, &req, sizeof(req), &resp, 686 sizeof(resp)); 687 if (err) { 688 netdev_err(ndev, "Failed to destroy WQ object: %d\n", err); 689 return; 690 } 691 692 err = mana_verify_resp_hdr(&resp.hdr, MANA_DESTROY_WQ_OBJ, 693 sizeof(resp)); 694 if (err || resp.hdr.status) 695 netdev_err(ndev, "Failed to destroy WQ object: %d, 0x%x\n", err, 696 resp.hdr.status); 697 } 698 699 static void mana_destroy_eq(struct mana_context *ac) 700 { 701 struct gdma_context *gc = ac->gdma_dev->gdma_context; 702 struct gdma_queue *eq; 703 int i; 704 705 if (!ac->eqs) 706 return; 707 708 for (i = 0; i < gc->max_num_queues; i++) { 709 eq = ac->eqs[i].eq; 710 if (!eq) 711 continue; 712 713 mana_gd_destroy_queue(gc, eq); 714 } 715 716 kfree(ac->eqs); 717 ac->eqs = NULL; 718 } 719 720 static int mana_create_eq(struct mana_context *ac) 721 { 722 struct gdma_dev *gd = ac->gdma_dev; 723 struct gdma_context *gc = gd->gdma_context; 724 struct gdma_queue_spec spec = {}; 725 int err; 726 int i; 727 728 ac->eqs = kcalloc(gc->max_num_queues, sizeof(struct mana_eq), 729 GFP_KERNEL); 730 if (!ac->eqs) 731 return -ENOMEM; 732 733 spec.type = GDMA_EQ; 734 spec.monitor_avl_buf = false; 735 spec.queue_size = EQ_SIZE; 736 spec.eq.callback = NULL; 737 spec.eq.context = ac->eqs; 738 spec.eq.log2_throttle_limit = LOG2_EQ_THROTTLE; 739 740 for (i = 0; i < gc->max_num_queues; i++) { 741 err = mana_gd_create_mana_eq(gd, &spec, &ac->eqs[i].eq); 742 if (err) 743 goto out; 744 } 745 746 return 0; 747 out: 748 mana_destroy_eq(ac); 749 return err; 750 } 751 752 static int mana_move_wq_tail(struct gdma_queue *wq, u32 num_units) 753 { 754 u32 used_space_old; 755 u32 used_space_new; 756 757 used_space_old = wq->head - wq->tail; 758 used_space_new = wq->head - (wq->tail + num_units); 759 760 if (WARN_ON_ONCE(used_space_new > used_space_old)) 761 return -ERANGE; 762 763 wq->tail += num_units; 764 return 0; 765 } 766 767 static void mana_unmap_skb(struct sk_buff *skb, struct mana_port_context *apc) 768 { 769 struct mana_skb_head *ash = (struct mana_skb_head *)skb->head; 770 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 771 struct device *dev = gc->dev; 772 int i; 773 774 dma_unmap_single(dev, ash->dma_handle[0], ash->size[0], DMA_TO_DEVICE); 775 776 for (i = 1; i < skb_shinfo(skb)->nr_frags + 1; i++) 777 dma_unmap_page(dev, ash->dma_handle[i], ash->size[i], 778 DMA_TO_DEVICE); 779 } 780 781 static void mana_poll_tx_cq(struct mana_cq *cq) 782 { 783 struct gdma_comp *completions = cq->gdma_comp_buf; 784 struct gdma_posted_wqe_info *wqe_info; 785 unsigned int pkt_transmitted = 0; 786 unsigned int wqe_unit_cnt = 0; 787 struct mana_txq *txq = cq->txq; 788 struct mana_port_context *apc; 789 struct netdev_queue *net_txq; 790 struct gdma_queue *gdma_wq; 791 unsigned int avail_space; 792 struct net_device *ndev; 793 struct sk_buff *skb; 794 bool txq_stopped; 795 int comp_read; 796 int i; 797 798 ndev = txq->ndev; 799 apc = netdev_priv(ndev); 800 801 comp_read = mana_gd_poll_cq(cq->gdma_cq, completions, 802 CQE_POLLING_BUFFER); 803 804 if (comp_read < 1) 805 return; 806 807 for (i = 0; i < comp_read; i++) { 808 struct mana_tx_comp_oob *cqe_oob; 809 810 if (WARN_ON_ONCE(!completions[i].is_sq)) 811 return; 812 813 cqe_oob = (struct mana_tx_comp_oob *)completions[i].cqe_data; 814 if (WARN_ON_ONCE(cqe_oob->cqe_hdr.client_type != 815 MANA_CQE_COMPLETION)) 816 return; 817 818 switch (cqe_oob->cqe_hdr.cqe_type) { 819 case CQE_TX_OKAY: 820 break; 821 822 case CQE_TX_SA_DROP: 823 case CQE_TX_MTU_DROP: 824 case CQE_TX_INVALID_OOB: 825 case CQE_TX_INVALID_ETH_TYPE: 826 case CQE_TX_HDR_PROCESSING_ERROR: 827 case CQE_TX_VF_DISABLED: 828 case CQE_TX_VPORT_IDX_OUT_OF_RANGE: 829 case CQE_TX_VPORT_DISABLED: 830 case CQE_TX_VLAN_TAGGING_VIOLATION: 831 WARN_ONCE(1, "TX: CQE error %d: ignored.\n", 832 cqe_oob->cqe_hdr.cqe_type); 833 break; 834 835 default: 836 /* If the CQE type is unexpected, log an error, assert, 837 * and go through the error path. 838 */ 839 WARN_ONCE(1, "TX: Unexpected CQE type %d: HW BUG?\n", 840 cqe_oob->cqe_hdr.cqe_type); 841 return; 842 } 843 844 if (WARN_ON_ONCE(txq->gdma_txq_id != completions[i].wq_num)) 845 return; 846 847 skb = skb_dequeue(&txq->pending_skbs); 848 if (WARN_ON_ONCE(!skb)) 849 return; 850 851 wqe_info = (struct gdma_posted_wqe_info *)skb->cb; 852 wqe_unit_cnt += wqe_info->wqe_size_in_bu; 853 854 mana_unmap_skb(skb, apc); 855 856 napi_consume_skb(skb, cq->budget); 857 858 pkt_transmitted++; 859 } 860 861 if (WARN_ON_ONCE(wqe_unit_cnt == 0)) 862 return; 863 864 mana_move_wq_tail(txq->gdma_sq, wqe_unit_cnt); 865 866 gdma_wq = txq->gdma_sq; 867 avail_space = mana_gd_wq_avail_space(gdma_wq); 868 869 /* Ensure tail updated before checking q stop */ 870 smp_mb(); 871 872 net_txq = txq->net_txq; 873 txq_stopped = netif_tx_queue_stopped(net_txq); 874 875 /* Ensure checking txq_stopped before apc->port_is_up. */ 876 smp_rmb(); 877 878 if (txq_stopped && apc->port_is_up && avail_space >= MAX_TX_WQE_SIZE) { 879 netif_tx_wake_queue(net_txq); 880 apc->eth_stats.wake_queue++; 881 } 882 883 if (atomic_sub_return(pkt_transmitted, &txq->pending_sends) < 0) 884 WARN_ON_ONCE(1); 885 886 cq->work_done = pkt_transmitted; 887 } 888 889 static void mana_post_pkt_rxq(struct mana_rxq *rxq) 890 { 891 struct mana_recv_buf_oob *recv_buf_oob; 892 u32 curr_index; 893 int err; 894 895 curr_index = rxq->buf_index++; 896 if (rxq->buf_index == rxq->num_rx_buf) 897 rxq->buf_index = 0; 898 899 recv_buf_oob = &rxq->rx_oobs[curr_index]; 900 901 err = mana_gd_post_and_ring(rxq->gdma_rq, &recv_buf_oob->wqe_req, 902 &recv_buf_oob->wqe_inf); 903 if (WARN_ON_ONCE(err)) 904 return; 905 906 WARN_ON_ONCE(recv_buf_oob->wqe_inf.wqe_size_in_bu != 1); 907 } 908 909 static void mana_rx_skb(void *buf_va, struct mana_rxcomp_oob *cqe, 910 struct mana_rxq *rxq) 911 { 912 struct mana_stats *rx_stats = &rxq->stats; 913 struct net_device *ndev = rxq->ndev; 914 uint pkt_len = cqe->ppi[0].pkt_len; 915 u16 rxq_idx = rxq->rxq_idx; 916 struct napi_struct *napi; 917 struct sk_buff *skb; 918 u32 hash_value; 919 920 rxq->rx_cq.work_done++; 921 napi = &rxq->rx_cq.napi; 922 923 if (!buf_va) { 924 ++ndev->stats.rx_dropped; 925 return; 926 } 927 928 skb = build_skb(buf_va, PAGE_SIZE); 929 930 if (!skb) { 931 free_page((unsigned long)buf_va); 932 ++ndev->stats.rx_dropped; 933 return; 934 } 935 936 skb_put(skb, pkt_len); 937 skb->dev = napi->dev; 938 939 skb->protocol = eth_type_trans(skb, ndev); 940 skb_checksum_none_assert(skb); 941 skb_record_rx_queue(skb, rxq_idx); 942 943 if ((ndev->features & NETIF_F_RXCSUM) && cqe->rx_iphdr_csum_succeed) { 944 if (cqe->rx_tcp_csum_succeed || cqe->rx_udp_csum_succeed) 945 skb->ip_summed = CHECKSUM_UNNECESSARY; 946 } 947 948 if (cqe->rx_hashtype != 0 && (ndev->features & NETIF_F_RXHASH)) { 949 hash_value = cqe->ppi[0].pkt_hash; 950 951 if (cqe->rx_hashtype & MANA_HASH_L4) 952 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L4); 953 else 954 skb_set_hash(skb, hash_value, PKT_HASH_TYPE_L3); 955 } 956 957 napi_gro_receive(napi, skb); 958 959 u64_stats_update_begin(&rx_stats->syncp); 960 rx_stats->packets++; 961 rx_stats->bytes += pkt_len; 962 u64_stats_update_end(&rx_stats->syncp); 963 } 964 965 static void mana_process_rx_cqe(struct mana_rxq *rxq, struct mana_cq *cq, 966 struct gdma_comp *cqe) 967 { 968 struct mana_rxcomp_oob *oob = (struct mana_rxcomp_oob *)cqe->cqe_data; 969 struct gdma_context *gc = rxq->gdma_rq->gdma_dev->gdma_context; 970 struct net_device *ndev = rxq->ndev; 971 struct mana_recv_buf_oob *rxbuf_oob; 972 struct device *dev = gc->dev; 973 void *new_buf, *old_buf; 974 struct page *new_page; 975 u32 curr, pktlen; 976 dma_addr_t da; 977 978 switch (oob->cqe_hdr.cqe_type) { 979 case CQE_RX_OKAY: 980 break; 981 982 case CQE_RX_TRUNCATED: 983 netdev_err(ndev, "Dropped a truncated packet\n"); 984 return; 985 986 case CQE_RX_COALESCED_4: 987 netdev_err(ndev, "RX coalescing is unsupported\n"); 988 return; 989 990 case CQE_RX_OBJECT_FENCE: 991 netdev_err(ndev, "RX Fencing is unsupported\n"); 992 return; 993 994 default: 995 netdev_err(ndev, "Unknown RX CQE type = %d\n", 996 oob->cqe_hdr.cqe_type); 997 return; 998 } 999 1000 if (oob->cqe_hdr.cqe_type != CQE_RX_OKAY) 1001 return; 1002 1003 pktlen = oob->ppi[0].pkt_len; 1004 1005 if (pktlen == 0) { 1006 /* data packets should never have packetlength of zero */ 1007 netdev_err(ndev, "RX pkt len=0, rq=%u, cq=%u, rxobj=0x%llx\n", 1008 rxq->gdma_id, cq->gdma_id, rxq->rxobj); 1009 return; 1010 } 1011 1012 curr = rxq->buf_index; 1013 rxbuf_oob = &rxq->rx_oobs[curr]; 1014 WARN_ON_ONCE(rxbuf_oob->wqe_inf.wqe_size_in_bu != 1); 1015 1016 new_page = alloc_page(GFP_ATOMIC); 1017 1018 if (new_page) { 1019 da = dma_map_page(dev, new_page, 0, rxq->datasize, 1020 DMA_FROM_DEVICE); 1021 1022 if (dma_mapping_error(dev, da)) { 1023 __free_page(new_page); 1024 new_page = NULL; 1025 } 1026 } 1027 1028 new_buf = new_page ? page_to_virt(new_page) : NULL; 1029 1030 if (new_buf) { 1031 dma_unmap_page(dev, rxbuf_oob->buf_dma_addr, rxq->datasize, 1032 DMA_FROM_DEVICE); 1033 1034 old_buf = rxbuf_oob->buf_va; 1035 1036 /* refresh the rxbuf_oob with the new page */ 1037 rxbuf_oob->buf_va = new_buf; 1038 rxbuf_oob->buf_dma_addr = da; 1039 rxbuf_oob->sgl[0].address = rxbuf_oob->buf_dma_addr; 1040 } else { 1041 old_buf = NULL; /* drop the packet if no memory */ 1042 } 1043 1044 mana_rx_skb(old_buf, oob, rxq); 1045 1046 mana_move_wq_tail(rxq->gdma_rq, rxbuf_oob->wqe_inf.wqe_size_in_bu); 1047 1048 mana_post_pkt_rxq(rxq); 1049 } 1050 1051 static void mana_poll_rx_cq(struct mana_cq *cq) 1052 { 1053 struct gdma_comp *comp = cq->gdma_comp_buf; 1054 int comp_read, i; 1055 1056 comp_read = mana_gd_poll_cq(cq->gdma_cq, comp, CQE_POLLING_BUFFER); 1057 WARN_ON_ONCE(comp_read > CQE_POLLING_BUFFER); 1058 1059 for (i = 0; i < comp_read; i++) { 1060 if (WARN_ON_ONCE(comp[i].is_sq)) 1061 return; 1062 1063 /* verify recv cqe references the right rxq */ 1064 if (WARN_ON_ONCE(comp[i].wq_num != cq->rxq->gdma_id)) 1065 return; 1066 1067 mana_process_rx_cqe(cq->rxq, cq, &comp[i]); 1068 } 1069 } 1070 1071 static void mana_cq_handler(void *context, struct gdma_queue *gdma_queue) 1072 { 1073 struct mana_cq *cq = context; 1074 u8 arm_bit; 1075 1076 WARN_ON_ONCE(cq->gdma_cq != gdma_queue); 1077 1078 if (cq->type == MANA_CQ_TYPE_RX) 1079 mana_poll_rx_cq(cq); 1080 else 1081 mana_poll_tx_cq(cq); 1082 1083 if (cq->work_done < cq->budget && 1084 napi_complete_done(&cq->napi, cq->work_done)) { 1085 arm_bit = SET_ARM_BIT; 1086 } else { 1087 arm_bit = 0; 1088 } 1089 1090 mana_gd_ring_cq(gdma_queue, arm_bit); 1091 } 1092 1093 static int mana_poll(struct napi_struct *napi, int budget) 1094 { 1095 struct mana_cq *cq = container_of(napi, struct mana_cq, napi); 1096 1097 cq->work_done = 0; 1098 cq->budget = budget; 1099 1100 mana_cq_handler(cq, cq->gdma_cq); 1101 1102 return min(cq->work_done, budget); 1103 } 1104 1105 static void mana_schedule_napi(void *context, struct gdma_queue *gdma_queue) 1106 { 1107 struct mana_cq *cq = context; 1108 1109 napi_schedule_irqoff(&cq->napi); 1110 } 1111 1112 static void mana_deinit_cq(struct mana_port_context *apc, struct mana_cq *cq) 1113 { 1114 struct gdma_dev *gd = apc->ac->gdma_dev; 1115 1116 if (!cq->gdma_cq) 1117 return; 1118 1119 mana_gd_destroy_queue(gd->gdma_context, cq->gdma_cq); 1120 } 1121 1122 static void mana_deinit_txq(struct mana_port_context *apc, struct mana_txq *txq) 1123 { 1124 struct gdma_dev *gd = apc->ac->gdma_dev; 1125 1126 if (!txq->gdma_sq) 1127 return; 1128 1129 mana_gd_destroy_queue(gd->gdma_context, txq->gdma_sq); 1130 } 1131 1132 static void mana_destroy_txq(struct mana_port_context *apc) 1133 { 1134 struct napi_struct *napi; 1135 int i; 1136 1137 if (!apc->tx_qp) 1138 return; 1139 1140 for (i = 0; i < apc->num_queues; i++) { 1141 napi = &apc->tx_qp[i].tx_cq.napi; 1142 napi_synchronize(napi); 1143 napi_disable(napi); 1144 netif_napi_del(napi); 1145 1146 mana_destroy_wq_obj(apc, GDMA_SQ, apc->tx_qp[i].tx_object); 1147 1148 mana_deinit_cq(apc, &apc->tx_qp[i].tx_cq); 1149 1150 mana_deinit_txq(apc, &apc->tx_qp[i].txq); 1151 } 1152 1153 kfree(apc->tx_qp); 1154 apc->tx_qp = NULL; 1155 } 1156 1157 static int mana_create_txq(struct mana_port_context *apc, 1158 struct net_device *net) 1159 { 1160 struct mana_context *ac = apc->ac; 1161 struct gdma_dev *gd = ac->gdma_dev; 1162 struct mana_obj_spec wq_spec; 1163 struct mana_obj_spec cq_spec; 1164 struct gdma_queue_spec spec; 1165 struct gdma_context *gc; 1166 struct mana_txq *txq; 1167 struct mana_cq *cq; 1168 u32 txq_size; 1169 u32 cq_size; 1170 int err; 1171 int i; 1172 1173 apc->tx_qp = kcalloc(apc->num_queues, sizeof(struct mana_tx_qp), 1174 GFP_KERNEL); 1175 if (!apc->tx_qp) 1176 return -ENOMEM; 1177 1178 /* The minimum size of the WQE is 32 bytes, hence 1179 * MAX_SEND_BUFFERS_PER_QUEUE represents the maximum number of WQEs 1180 * the SQ can store. This value is then used to size other queues 1181 * to prevent overflow. 1182 */ 1183 txq_size = MAX_SEND_BUFFERS_PER_QUEUE * 32; 1184 BUILD_BUG_ON(!PAGE_ALIGNED(txq_size)); 1185 1186 cq_size = MAX_SEND_BUFFERS_PER_QUEUE * COMP_ENTRY_SIZE; 1187 cq_size = PAGE_ALIGN(cq_size); 1188 1189 gc = gd->gdma_context; 1190 1191 for (i = 0; i < apc->num_queues; i++) { 1192 apc->tx_qp[i].tx_object = INVALID_MANA_HANDLE; 1193 1194 /* Create SQ */ 1195 txq = &apc->tx_qp[i].txq; 1196 1197 u64_stats_init(&txq->stats.syncp); 1198 txq->ndev = net; 1199 txq->net_txq = netdev_get_tx_queue(net, i); 1200 txq->vp_offset = apc->tx_vp_offset; 1201 skb_queue_head_init(&txq->pending_skbs); 1202 1203 memset(&spec, 0, sizeof(spec)); 1204 spec.type = GDMA_SQ; 1205 spec.monitor_avl_buf = true; 1206 spec.queue_size = txq_size; 1207 err = mana_gd_create_mana_wq_cq(gd, &spec, &txq->gdma_sq); 1208 if (err) 1209 goto out; 1210 1211 /* Create SQ's CQ */ 1212 cq = &apc->tx_qp[i].tx_cq; 1213 cq->type = MANA_CQ_TYPE_TX; 1214 1215 cq->txq = txq; 1216 1217 memset(&spec, 0, sizeof(spec)); 1218 spec.type = GDMA_CQ; 1219 spec.monitor_avl_buf = false; 1220 spec.queue_size = cq_size; 1221 spec.cq.callback = mana_schedule_napi; 1222 spec.cq.parent_eq = ac->eqs[i].eq; 1223 spec.cq.context = cq; 1224 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1225 if (err) 1226 goto out; 1227 1228 memset(&wq_spec, 0, sizeof(wq_spec)); 1229 memset(&cq_spec, 0, sizeof(cq_spec)); 1230 1231 wq_spec.gdma_region = txq->gdma_sq->mem_info.gdma_region; 1232 wq_spec.queue_size = txq->gdma_sq->queue_size; 1233 1234 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 1235 cq_spec.queue_size = cq->gdma_cq->queue_size; 1236 cq_spec.modr_ctx_id = 0; 1237 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1238 1239 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_SQ, 1240 &wq_spec, &cq_spec, 1241 &apc->tx_qp[i].tx_object); 1242 1243 if (err) 1244 goto out; 1245 1246 txq->gdma_sq->id = wq_spec.queue_index; 1247 cq->gdma_cq->id = cq_spec.queue_index; 1248 1249 txq->gdma_sq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1250 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1251 1252 txq->gdma_txq_id = txq->gdma_sq->id; 1253 1254 cq->gdma_id = cq->gdma_cq->id; 1255 1256 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1257 err = -EINVAL; 1258 goto out; 1259 } 1260 1261 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1262 1263 netif_tx_napi_add(net, &cq->napi, mana_poll, NAPI_POLL_WEIGHT); 1264 napi_enable(&cq->napi); 1265 1266 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1267 } 1268 1269 return 0; 1270 out: 1271 mana_destroy_txq(apc); 1272 return err; 1273 } 1274 1275 static void mana_destroy_rxq(struct mana_port_context *apc, 1276 struct mana_rxq *rxq, bool validate_state) 1277 1278 { 1279 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1280 struct mana_recv_buf_oob *rx_oob; 1281 struct device *dev = gc->dev; 1282 struct napi_struct *napi; 1283 int i; 1284 1285 if (!rxq) 1286 return; 1287 1288 napi = &rxq->rx_cq.napi; 1289 1290 if (validate_state) 1291 napi_synchronize(napi); 1292 1293 napi_disable(napi); 1294 netif_napi_del(napi); 1295 1296 mana_destroy_wq_obj(apc, GDMA_RQ, rxq->rxobj); 1297 1298 mana_deinit_cq(apc, &rxq->rx_cq); 1299 1300 for (i = 0; i < rxq->num_rx_buf; i++) { 1301 rx_oob = &rxq->rx_oobs[i]; 1302 1303 if (!rx_oob->buf_va) 1304 continue; 1305 1306 dma_unmap_page(dev, rx_oob->buf_dma_addr, rxq->datasize, 1307 DMA_FROM_DEVICE); 1308 1309 free_page((unsigned long)rx_oob->buf_va); 1310 rx_oob->buf_va = NULL; 1311 } 1312 1313 if (rxq->gdma_rq) 1314 mana_gd_destroy_queue(gc, rxq->gdma_rq); 1315 1316 kfree(rxq); 1317 } 1318 1319 #define MANA_WQE_HEADER_SIZE 16 1320 #define MANA_WQE_SGE_SIZE 16 1321 1322 static int mana_alloc_rx_wqe(struct mana_port_context *apc, 1323 struct mana_rxq *rxq, u32 *rxq_size, u32 *cq_size) 1324 { 1325 struct gdma_context *gc = apc->ac->gdma_dev->gdma_context; 1326 struct mana_recv_buf_oob *rx_oob; 1327 struct device *dev = gc->dev; 1328 struct page *page; 1329 dma_addr_t da; 1330 u32 buf_idx; 1331 1332 WARN_ON(rxq->datasize == 0 || rxq->datasize > PAGE_SIZE); 1333 1334 *rxq_size = 0; 1335 *cq_size = 0; 1336 1337 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1338 rx_oob = &rxq->rx_oobs[buf_idx]; 1339 memset(rx_oob, 0, sizeof(*rx_oob)); 1340 1341 page = alloc_page(GFP_KERNEL); 1342 if (!page) 1343 return -ENOMEM; 1344 1345 da = dma_map_page(dev, page, 0, rxq->datasize, DMA_FROM_DEVICE); 1346 1347 if (dma_mapping_error(dev, da)) { 1348 __free_page(page); 1349 return -ENOMEM; 1350 } 1351 1352 rx_oob->buf_va = page_to_virt(page); 1353 rx_oob->buf_dma_addr = da; 1354 1355 rx_oob->num_sge = 1; 1356 rx_oob->sgl[0].address = rx_oob->buf_dma_addr; 1357 rx_oob->sgl[0].size = rxq->datasize; 1358 rx_oob->sgl[0].mem_key = apc->ac->gdma_dev->gpa_mkey; 1359 1360 rx_oob->wqe_req.sgl = rx_oob->sgl; 1361 rx_oob->wqe_req.num_sge = rx_oob->num_sge; 1362 rx_oob->wqe_req.inline_oob_size = 0; 1363 rx_oob->wqe_req.inline_oob_data = NULL; 1364 rx_oob->wqe_req.flags = 0; 1365 rx_oob->wqe_req.client_data_unit = 0; 1366 1367 *rxq_size += ALIGN(MANA_WQE_HEADER_SIZE + 1368 MANA_WQE_SGE_SIZE * rx_oob->num_sge, 32); 1369 *cq_size += COMP_ENTRY_SIZE; 1370 } 1371 1372 return 0; 1373 } 1374 1375 static int mana_push_wqe(struct mana_rxq *rxq) 1376 { 1377 struct mana_recv_buf_oob *rx_oob; 1378 u32 buf_idx; 1379 int err; 1380 1381 for (buf_idx = 0; buf_idx < rxq->num_rx_buf; buf_idx++) { 1382 rx_oob = &rxq->rx_oobs[buf_idx]; 1383 1384 err = mana_gd_post_and_ring(rxq->gdma_rq, &rx_oob->wqe_req, 1385 &rx_oob->wqe_inf); 1386 if (err) 1387 return -ENOSPC; 1388 } 1389 1390 return 0; 1391 } 1392 1393 static struct mana_rxq *mana_create_rxq(struct mana_port_context *apc, 1394 u32 rxq_idx, struct mana_eq *eq, 1395 struct net_device *ndev) 1396 { 1397 struct gdma_dev *gd = apc->ac->gdma_dev; 1398 struct mana_obj_spec wq_spec; 1399 struct mana_obj_spec cq_spec; 1400 struct gdma_queue_spec spec; 1401 struct mana_cq *cq = NULL; 1402 struct gdma_context *gc; 1403 u32 cq_size, rq_size; 1404 struct mana_rxq *rxq; 1405 int err; 1406 1407 gc = gd->gdma_context; 1408 1409 rxq = kzalloc(struct_size(rxq, rx_oobs, RX_BUFFERS_PER_QUEUE), 1410 GFP_KERNEL); 1411 if (!rxq) 1412 return NULL; 1413 1414 rxq->ndev = ndev; 1415 rxq->num_rx_buf = RX_BUFFERS_PER_QUEUE; 1416 rxq->rxq_idx = rxq_idx; 1417 rxq->datasize = ALIGN(MAX_FRAME_SIZE, 64); 1418 rxq->rxobj = INVALID_MANA_HANDLE; 1419 1420 err = mana_alloc_rx_wqe(apc, rxq, &rq_size, &cq_size); 1421 if (err) 1422 goto out; 1423 1424 rq_size = PAGE_ALIGN(rq_size); 1425 cq_size = PAGE_ALIGN(cq_size); 1426 1427 /* Create RQ */ 1428 memset(&spec, 0, sizeof(spec)); 1429 spec.type = GDMA_RQ; 1430 spec.monitor_avl_buf = true; 1431 spec.queue_size = rq_size; 1432 err = mana_gd_create_mana_wq_cq(gd, &spec, &rxq->gdma_rq); 1433 if (err) 1434 goto out; 1435 1436 /* Create RQ's CQ */ 1437 cq = &rxq->rx_cq; 1438 cq->type = MANA_CQ_TYPE_RX; 1439 cq->rxq = rxq; 1440 1441 memset(&spec, 0, sizeof(spec)); 1442 spec.type = GDMA_CQ; 1443 spec.monitor_avl_buf = false; 1444 spec.queue_size = cq_size; 1445 spec.cq.callback = mana_schedule_napi; 1446 spec.cq.parent_eq = eq->eq; 1447 spec.cq.context = cq; 1448 err = mana_gd_create_mana_wq_cq(gd, &spec, &cq->gdma_cq); 1449 if (err) 1450 goto out; 1451 1452 memset(&wq_spec, 0, sizeof(wq_spec)); 1453 memset(&cq_spec, 0, sizeof(cq_spec)); 1454 wq_spec.gdma_region = rxq->gdma_rq->mem_info.gdma_region; 1455 wq_spec.queue_size = rxq->gdma_rq->queue_size; 1456 1457 cq_spec.gdma_region = cq->gdma_cq->mem_info.gdma_region; 1458 cq_spec.queue_size = cq->gdma_cq->queue_size; 1459 cq_spec.modr_ctx_id = 0; 1460 cq_spec.attached_eq = cq->gdma_cq->cq.parent->id; 1461 1462 err = mana_create_wq_obj(apc, apc->port_handle, GDMA_RQ, 1463 &wq_spec, &cq_spec, &rxq->rxobj); 1464 if (err) 1465 goto out; 1466 1467 rxq->gdma_rq->id = wq_spec.queue_index; 1468 cq->gdma_cq->id = cq_spec.queue_index; 1469 1470 rxq->gdma_rq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1471 cq->gdma_cq->mem_info.gdma_region = GDMA_INVALID_DMA_REGION; 1472 1473 rxq->gdma_id = rxq->gdma_rq->id; 1474 cq->gdma_id = cq->gdma_cq->id; 1475 1476 err = mana_push_wqe(rxq); 1477 if (err) 1478 goto out; 1479 1480 if (WARN_ON(cq->gdma_id >= gc->max_num_cqs)) { 1481 err = -EINVAL; 1482 goto out; 1483 } 1484 1485 gc->cq_table[cq->gdma_id] = cq->gdma_cq; 1486 1487 netif_napi_add(ndev, &cq->napi, mana_poll, 1); 1488 napi_enable(&cq->napi); 1489 1490 mana_gd_ring_cq(cq->gdma_cq, SET_ARM_BIT); 1491 out: 1492 if (!err) 1493 return rxq; 1494 1495 netdev_err(ndev, "Failed to create RXQ: err = %d\n", err); 1496 1497 mana_destroy_rxq(apc, rxq, false); 1498 1499 if (cq) 1500 mana_deinit_cq(apc, cq); 1501 1502 return NULL; 1503 } 1504 1505 static int mana_add_rx_queues(struct mana_port_context *apc, 1506 struct net_device *ndev) 1507 { 1508 struct mana_context *ac = apc->ac; 1509 struct mana_rxq *rxq; 1510 int err = 0; 1511 int i; 1512 1513 for (i = 0; i < apc->num_queues; i++) { 1514 rxq = mana_create_rxq(apc, i, &ac->eqs[i], ndev); 1515 if (!rxq) { 1516 err = -ENOMEM; 1517 goto out; 1518 } 1519 1520 u64_stats_init(&rxq->stats.syncp); 1521 1522 apc->rxqs[i] = rxq; 1523 } 1524 1525 apc->default_rxobj = apc->rxqs[0]->rxobj; 1526 out: 1527 return err; 1528 } 1529 1530 static void mana_destroy_vport(struct mana_port_context *apc) 1531 { 1532 struct mana_rxq *rxq; 1533 u32 rxq_idx; 1534 1535 for (rxq_idx = 0; rxq_idx < apc->num_queues; rxq_idx++) { 1536 rxq = apc->rxqs[rxq_idx]; 1537 if (!rxq) 1538 continue; 1539 1540 mana_destroy_rxq(apc, rxq, true); 1541 apc->rxqs[rxq_idx] = NULL; 1542 } 1543 1544 mana_destroy_txq(apc); 1545 } 1546 1547 static int mana_create_vport(struct mana_port_context *apc, 1548 struct net_device *net) 1549 { 1550 struct gdma_dev *gd = apc->ac->gdma_dev; 1551 int err; 1552 1553 apc->default_rxobj = INVALID_MANA_HANDLE; 1554 1555 err = mana_cfg_vport(apc, gd->pdid, gd->doorbell); 1556 if (err) 1557 return err; 1558 1559 return mana_create_txq(apc, net); 1560 } 1561 1562 static void mana_rss_table_init(struct mana_port_context *apc) 1563 { 1564 int i; 1565 1566 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) 1567 apc->indir_table[i] = 1568 ethtool_rxfh_indir_default(i, apc->num_queues); 1569 } 1570 1571 int mana_config_rss(struct mana_port_context *apc, enum TRI_STATE rx, 1572 bool update_hash, bool update_tab) 1573 { 1574 u32 queue_idx; 1575 int i; 1576 1577 if (update_tab) { 1578 for (i = 0; i < MANA_INDIRECT_TABLE_SIZE; i++) { 1579 queue_idx = apc->indir_table[i]; 1580 apc->rxobj_table[i] = apc->rxqs[queue_idx]->rxobj; 1581 } 1582 } 1583 1584 return mana_cfg_vport_steering(apc, rx, true, update_hash, update_tab); 1585 } 1586 1587 static int mana_init_port(struct net_device *ndev) 1588 { 1589 struct mana_port_context *apc = netdev_priv(ndev); 1590 u32 max_txq, max_rxq, max_queues; 1591 int port_idx = apc->port_idx; 1592 u32 num_indirect_entries; 1593 int err; 1594 1595 err = mana_init_port_context(apc); 1596 if (err) 1597 return err; 1598 1599 err = mana_query_vport_cfg(apc, port_idx, &max_txq, &max_rxq, 1600 &num_indirect_entries); 1601 if (err) { 1602 netdev_err(ndev, "Failed to query info for vPort %d\n", 1603 port_idx); 1604 goto reset_apc; 1605 } 1606 1607 max_queues = min_t(u32, max_txq, max_rxq); 1608 if (apc->max_queues > max_queues) 1609 apc->max_queues = max_queues; 1610 1611 if (apc->num_queues > apc->max_queues) 1612 apc->num_queues = apc->max_queues; 1613 1614 eth_hw_addr_set(ndev, apc->mac_addr); 1615 1616 return 0; 1617 1618 reset_apc: 1619 kfree(apc->rxqs); 1620 apc->rxqs = NULL; 1621 return err; 1622 } 1623 1624 int mana_alloc_queues(struct net_device *ndev) 1625 { 1626 struct mana_port_context *apc = netdev_priv(ndev); 1627 int err; 1628 1629 err = mana_create_vport(apc, ndev); 1630 if (err) 1631 return err; 1632 1633 err = netif_set_real_num_tx_queues(ndev, apc->num_queues); 1634 if (err) 1635 goto destroy_vport; 1636 1637 err = mana_add_rx_queues(apc, ndev); 1638 if (err) 1639 goto destroy_vport; 1640 1641 apc->rss_state = apc->num_queues > 1 ? TRI_STATE_TRUE : TRI_STATE_FALSE; 1642 1643 err = netif_set_real_num_rx_queues(ndev, apc->num_queues); 1644 if (err) 1645 goto destroy_vport; 1646 1647 mana_rss_table_init(apc); 1648 1649 err = mana_config_rss(apc, TRI_STATE_TRUE, true, true); 1650 if (err) 1651 goto destroy_vport; 1652 1653 return 0; 1654 1655 destroy_vport: 1656 mana_destroy_vport(apc); 1657 return err; 1658 } 1659 1660 int mana_attach(struct net_device *ndev) 1661 { 1662 struct mana_port_context *apc = netdev_priv(ndev); 1663 int err; 1664 1665 ASSERT_RTNL(); 1666 1667 err = mana_init_port(ndev); 1668 if (err) 1669 return err; 1670 1671 if (apc->port_st_save) { 1672 err = mana_alloc_queues(ndev); 1673 if (err) { 1674 mana_cleanup_port_context(apc); 1675 return err; 1676 } 1677 } 1678 1679 apc->port_is_up = apc->port_st_save; 1680 1681 /* Ensure port state updated before txq state */ 1682 smp_wmb(); 1683 1684 if (apc->port_is_up) 1685 netif_carrier_on(ndev); 1686 1687 netif_device_attach(ndev); 1688 1689 return 0; 1690 } 1691 1692 static int mana_dealloc_queues(struct net_device *ndev) 1693 { 1694 struct mana_port_context *apc = netdev_priv(ndev); 1695 struct mana_txq *txq; 1696 int i, err; 1697 1698 if (apc->port_is_up) 1699 return -EINVAL; 1700 1701 /* No packet can be transmitted now since apc->port_is_up is false. 1702 * There is still a tiny chance that mana_poll_tx_cq() can re-enable 1703 * a txq because it may not timely see apc->port_is_up being cleared 1704 * to false, but it doesn't matter since mana_start_xmit() drops any 1705 * new packets due to apc->port_is_up being false. 1706 * 1707 * Drain all the in-flight TX packets 1708 */ 1709 for (i = 0; i < apc->num_queues; i++) { 1710 txq = &apc->tx_qp[i].txq; 1711 1712 while (atomic_read(&txq->pending_sends) > 0) 1713 usleep_range(1000, 2000); 1714 } 1715 1716 /* We're 100% sure the queues can no longer be woken up, because 1717 * we're sure now mana_poll_tx_cq() can't be running. 1718 */ 1719 1720 apc->rss_state = TRI_STATE_FALSE; 1721 err = mana_config_rss(apc, TRI_STATE_FALSE, false, false); 1722 if (err) { 1723 netdev_err(ndev, "Failed to disable vPort: %d\n", err); 1724 return err; 1725 } 1726 1727 /* TODO: Implement RX fencing */ 1728 ssleep(1); 1729 1730 mana_destroy_vport(apc); 1731 1732 return 0; 1733 } 1734 1735 int mana_detach(struct net_device *ndev, bool from_close) 1736 { 1737 struct mana_port_context *apc = netdev_priv(ndev); 1738 int err; 1739 1740 ASSERT_RTNL(); 1741 1742 apc->port_st_save = apc->port_is_up; 1743 apc->port_is_up = false; 1744 1745 /* Ensure port state updated before txq state */ 1746 smp_wmb(); 1747 1748 netif_tx_disable(ndev); 1749 netif_carrier_off(ndev); 1750 1751 if (apc->port_st_save) { 1752 err = mana_dealloc_queues(ndev); 1753 if (err) 1754 return err; 1755 } 1756 1757 if (!from_close) { 1758 netif_device_detach(ndev); 1759 mana_cleanup_port_context(apc); 1760 } 1761 1762 return 0; 1763 } 1764 1765 static int mana_probe_port(struct mana_context *ac, int port_idx, 1766 struct net_device **ndev_storage) 1767 { 1768 struct gdma_context *gc = ac->gdma_dev->gdma_context; 1769 struct mana_port_context *apc; 1770 struct net_device *ndev; 1771 int err; 1772 1773 ndev = alloc_etherdev_mq(sizeof(struct mana_port_context), 1774 gc->max_num_queues); 1775 if (!ndev) 1776 return -ENOMEM; 1777 1778 *ndev_storage = ndev; 1779 1780 apc = netdev_priv(ndev); 1781 apc->ac = ac; 1782 apc->ndev = ndev; 1783 apc->max_queues = gc->max_num_queues; 1784 apc->num_queues = gc->max_num_queues; 1785 apc->port_handle = INVALID_MANA_HANDLE; 1786 apc->port_idx = port_idx; 1787 1788 ndev->netdev_ops = &mana_devops; 1789 ndev->ethtool_ops = &mana_ethtool_ops; 1790 ndev->mtu = ETH_DATA_LEN; 1791 ndev->max_mtu = ndev->mtu; 1792 ndev->min_mtu = ndev->mtu; 1793 ndev->needed_headroom = MANA_HEADROOM; 1794 SET_NETDEV_DEV(ndev, gc->dev); 1795 1796 netif_carrier_off(ndev); 1797 1798 netdev_rss_key_fill(apc->hashkey, MANA_HASH_KEY_SIZE); 1799 1800 err = mana_init_port(ndev); 1801 if (err) 1802 goto free_net; 1803 1804 netdev_lockdep_set_classes(ndev); 1805 1806 ndev->hw_features = NETIF_F_SG | NETIF_F_IP_CSUM | NETIF_F_IPV6_CSUM; 1807 ndev->hw_features |= NETIF_F_RXCSUM; 1808 ndev->hw_features |= NETIF_F_TSO | NETIF_F_TSO6; 1809 ndev->hw_features |= NETIF_F_RXHASH; 1810 ndev->features = ndev->hw_features; 1811 ndev->vlan_features = 0; 1812 1813 err = register_netdev(ndev); 1814 if (err) { 1815 netdev_err(ndev, "Unable to register netdev.\n"); 1816 goto reset_apc; 1817 } 1818 1819 return 0; 1820 1821 reset_apc: 1822 kfree(apc->rxqs); 1823 apc->rxqs = NULL; 1824 free_net: 1825 *ndev_storage = NULL; 1826 netdev_err(ndev, "Failed to probe vPort %d: %d\n", port_idx, err); 1827 free_netdev(ndev); 1828 return err; 1829 } 1830 1831 int mana_probe(struct gdma_dev *gd, bool resuming) 1832 { 1833 struct gdma_context *gc = gd->gdma_context; 1834 struct mana_context *ac = gd->driver_data; 1835 struct device *dev = gc->dev; 1836 u16 num_ports = 0; 1837 int err; 1838 int i; 1839 1840 dev_info(dev, 1841 "Microsoft Azure Network Adapter protocol version: %d.%d.%d\n", 1842 MANA_MAJOR_VERSION, MANA_MINOR_VERSION, MANA_MICRO_VERSION); 1843 1844 err = mana_gd_register_device(gd); 1845 if (err) 1846 return err; 1847 1848 if (!resuming) { 1849 ac = kzalloc(sizeof(*ac), GFP_KERNEL); 1850 if (!ac) 1851 return -ENOMEM; 1852 1853 ac->gdma_dev = gd; 1854 gd->driver_data = ac; 1855 } 1856 1857 err = mana_create_eq(ac); 1858 if (err) 1859 goto out; 1860 1861 err = mana_query_device_cfg(ac, MANA_MAJOR_VERSION, MANA_MINOR_VERSION, 1862 MANA_MICRO_VERSION, &num_ports); 1863 if (err) 1864 goto out; 1865 1866 if (!resuming) { 1867 ac->num_ports = num_ports; 1868 } else { 1869 if (ac->num_ports != num_ports) { 1870 dev_err(dev, "The number of vPorts changed: %d->%d\n", 1871 ac->num_ports, num_ports); 1872 err = -EPROTO; 1873 goto out; 1874 } 1875 } 1876 1877 if (ac->num_ports == 0) 1878 dev_err(dev, "Failed to detect any vPort\n"); 1879 1880 if (ac->num_ports > MAX_PORTS_IN_MANA_DEV) 1881 ac->num_ports = MAX_PORTS_IN_MANA_DEV; 1882 1883 if (!resuming) { 1884 for (i = 0; i < ac->num_ports; i++) { 1885 err = mana_probe_port(ac, i, &ac->ports[i]); 1886 if (err) 1887 break; 1888 } 1889 } else { 1890 for (i = 0; i < ac->num_ports; i++) { 1891 rtnl_lock(); 1892 err = mana_attach(ac->ports[i]); 1893 rtnl_unlock(); 1894 if (err) 1895 break; 1896 } 1897 } 1898 out: 1899 if (err) 1900 mana_remove(gd, false); 1901 1902 return err; 1903 } 1904 1905 void mana_remove(struct gdma_dev *gd, bool suspending) 1906 { 1907 struct gdma_context *gc = gd->gdma_context; 1908 struct mana_context *ac = gd->driver_data; 1909 struct device *dev = gc->dev; 1910 struct net_device *ndev; 1911 int err; 1912 int i; 1913 1914 for (i = 0; i < ac->num_ports; i++) { 1915 ndev = ac->ports[i]; 1916 if (!ndev) { 1917 if (i == 0) 1918 dev_err(dev, "No net device to remove\n"); 1919 goto out; 1920 } 1921 1922 /* All cleanup actions should stay after rtnl_lock(), otherwise 1923 * other functions may access partially cleaned up data. 1924 */ 1925 rtnl_lock(); 1926 1927 err = mana_detach(ndev, false); 1928 if (err) 1929 netdev_err(ndev, "Failed to detach vPort %d: %d\n", 1930 i, err); 1931 1932 if (suspending) { 1933 /* No need to unregister the ndev. */ 1934 rtnl_unlock(); 1935 continue; 1936 } 1937 1938 unregister_netdevice(ndev); 1939 1940 rtnl_unlock(); 1941 1942 free_netdev(ndev); 1943 } 1944 1945 mana_destroy_eq(ac); 1946 1947 out: 1948 mana_gd_deregister_device(gd); 1949 1950 if (suspending) 1951 return; 1952 1953 gd->driver_data = NULL; 1954 gd->gdma_context = NULL; 1955 kfree(ac); 1956 } 1957